diff --git a/cel/cel_test.go b/cel/cel_test.go
index 05adeb2f..09e54751 100644
--- a/cel/cel_test.go
+++ b/cel/cel_test.go
@@ -1969,6 +1969,48 @@ func TestCostLimit(t *testing.T) {
 	}
 }
 
+func TestCostTrackingConsistentAcrossEvals(t *testing.T) {
+	env := testEnv(t,
+		Variable("val1", IntType),
+		Variable("val2", IntType),
+	)
+	ast, iss := env.Compile(`val1 + val2`)
+	if iss.Err() != nil {
+		t.Fatalf("env.Compile() failed: %v", iss.Err())
+	}
+	checkedAst, iss := env.Check(ast)
+	if iss.Err() != nil {
+		t.Fatalf("env.Check() failed: %v", iss.Err())
+	}
+	program, err := env.Program(checkedAst, CostTracking(nil))
+	if err != nil {
+		t.Fatalf("env.Program() failed: %v", err)
+	}
+	input := map[string]any{"val1": 1, "val2": 2}
+
+	_, det1, err := program.Eval(input)
+	if err != nil {
+		t.Fatalf("first Eval() failed: %v", err)
+	}
+	cost1 := det1.ActualCost()
+	if cost1 == nil {
+		t.Fatal("first Eval() returned nil ActualCost")
+	}
+
+	_, det2, err := program.Eval(input)
+	if err != nil {
+		t.Fatalf("second Eval() failed: %v", err)
+	}
+	cost2 := det2.ActualCost()
+	if cost2 == nil {
+		t.Fatal("second Eval() returned nil ActualCost")
+	}
+
+	if *cost1 != *cost2 {
+		t.Errorf("ActualCost mismatch across evaluations: first=%d, second=%d", *cost1, *cost2)
+	}
+}
+
 func TestPartialVars(t *testing.T) {
 	env := testEnv(t,
 		Variable("x", StringType),
diff --git a/cel/program.go b/cel/program.go
index 79df0374..1fb65877 100644
--- a/cel/program.go
+++ b/cel/program.go
@@ -261,8 +261,16 @@ func newProgram(e *Env, a *ast.AST, opts []ProgramOption) (Program, error) {
 		if p.costLimit != nil {
 			costOpts = append(costOpts, interpreter.CostTrackerLimit(*p.costLimit))
 		}
+		// Creating a new cost tracker for each evaluation causes significant work that
+		// needs to be repeated for each evaluation even though the cost tracker is
+		// mostly read-only once constructed. Therefore it gets constructed
+		// once now and later a cheap clone is used for each evaluation.
+		tracker, err := interpreter.NewCostTracker(p.callCostEstimator, costOpts...)
+		if err != nil {
+			return nil, fmt.Errorf("construct cost tracker: %w", err)
+		}
 		trackerFactory := func() (*interpreter.CostTracker, error) {
-			return interpreter.NewCostTracker(p.callCostEstimator, costOpts...)
+			return tracker.Clone()
 		}
 		var observers []interpreter.PlannerOption
 		if p.evalOpts&(OptExhaustiveEval|OptTrackState) != 0 {
diff --git a/interpreter/runtimecost.go b/interpreter/runtimecost.go
index 0233e9a2..68e43101 100644
--- a/interpreter/runtimecost.go
+++ b/interpreter/runtimecost.go
@@ -233,6 +233,19 @@ type CostTracker struct {
 	stack refValStack
 }
 
+// Clone makes a shallow copy of the tracker.
+// The different clones can be used independently from
+// each other.
+func (c *CostTracker) Clone() (*CostTracker, error) {
+	tracker := &CostTracker{
+		Estimator:           c.Estimator,
+		overloadTrackers:    c.overloadTrackers,
+		Limit:               c.Limit,
+		presenceTestHasCost: c.presenceTestHasCost,
+	}
+	return tracker, nil
+}
+
 // ActualCost returns the runtime cost
 func (c *CostTracker) ActualCost() uint64 {
 	return c.cost
diff --git a/interpreter/runtimecost_test.go b/interpreter/runtimecost_test.go
index a55db6f5..597d73ef 100644
--- a/interpreter/runtimecost_test.go
+++ b/interpreter/runtimecost_test.go
@@ -134,6 +134,10 @@ func computeCost(t *testing.T, expr string, vars []*decls.VariableDecl, ctx Acti
 	if err != nil {
 		t.Fatalf("NewCostTracker() failed: %v", err)
 	}
+	costTracker, err = costTracker.Clone()
+	if err != nil {
+		t.Fatalf("checker.Clone() failed: %v", err)
+	}
 	checked, errs := checker.Check(parsed, s, env)
 	if len(errs.GetErrors()) != 0 {
 		t.Fatalf(`Failed to check expression "%s", error: %v`, expr, errs.GetErrors())